System prompts are the standing instructions AI follows (like 'always be neutral'); user prompts are what you ask it right now. Understanding the difference matters because a system prompt can silently shape every answer you get, while you might think you're controlling the output through your questions alone.
A system prompt is a hidden instruction set that defines how an AI model behaves globally within a tool. Your workplace AI might have a system prompt like "You are a professional assistant designed to help with business documentation. Prioritize accuracy and legal compliance. Never provide legal advice." This constrains the entire model's behavior independently of what you ask.
User prompts are your direct requests: "Summarize this email" or "Draft a response to my manager's feedback." Your user prompts operate within the constraints established by the system prompt. If the system prompt forbids legal advice, your user prompt can't override that constraint.
System prompts function as the AI's operating instructions. They establish priorities, define ethical guardrails, and shape tone. When you use Claude through Anthropic's official interface versus Claude through a third-party tool, the system prompt differs. The third-party tool might inject instructions like "Emphasize this company's service offerings" or "Decline requests for competitor analysis." The underlying Claude model is identical, but system-level instructions change behavior.
In workplace contexts, you rarely see the system prompt. It's typically hidden from users. This creates an asymmetry: you don't know what global constraints or priorities are shaping the AI's responses to your prompts.
If you're using an HR-focused AI tool that has a system prompt emphasizing "employer interests," the tool might subtly deprioritize documentation protecting your career. You ask it to help document unfair treatment, but the system prompt has instructed the model to "remain balanced and avoid inflammatory language." The system prompt overrides your intent.
This doesn't mean the tool is lying to you. It means there's a layer of instruction above your request that you can't see and can't directly override through your user prompts. Recognizing this invisible layer is critical for understanding why an AI tool behaves differently than you expected.
You can infer what system-level constraints exist by testing boundaries. Ask the tool to do something clearly outside professional norms: "Write me an angry, accusatory email to my boss." If the AI refuses and suggests professional alternatives instead, a system prompt is likely constraining the output. That's often good (prevents genuinely harmful outputs) but limits what you can ask it to do.
Try more specific tests: "Summarize this without corporate jargon or professional softening." If the tool insists on adding professional framing despite your request, a system prompt is likely enforcing that tone. This matters if you need raw, unfiltered documentation rather than polished versions.
Many workplace AI tools don't disclose their system prompts. You don't know whether the tool has been instructed to:
When using a tool with suspected hidden system prompts, explicitly counter them with your user prompts. If you suspect the tool softens critical language, request: "Present this information in direct, unambiguous language without professional euphemisms." If you suspect employer-side bias, frame requests explicitly: "Document this from my perspective as the affected employee, not from a neutral management standpoint."
For critical documentation, shift to tools where system prompts are transparent or minimal. Claude's default behavior is relatively transparent—it follows your instructions unless they violate specific Anthropic guidelines (which are documented). ChatGPT has similar transparency. Using these directly, rather than through intermediary tools, gives you more predictable behavior.
Create a test prompt library. Ask a tool the same question multiple ways and compare consistency. Inconsistency reveals underlying constraints. If you ask for "a summary of this conflict" versus "analysis of this manager's behavior," do you get different emphasis or tone? That variation hints at system-level priorities.
Document tools that seem to soften your language, add professional framing you didn't request, or subtly reframe your concerns. These are red flags for system prompts working against your interests. For documentation protecting your career, that's unacceptable.
Try this: Take a difficult workplace situation you're documenting. Write a raw, direct version of what happened without professional softening. Paste it into your AI tool and ask: "Help me document this professionally while preserving the factual substance." Compare the output. Did the tool preserve your core claims while adding professional tone, or did it subtly reframe your concerns? This reveals whether the system prompt is neutral (just adding polish) or biased (reframing your narrative).
Peri can explain this concept, give practical examples, help you decide whether it applies to your situation, or recommend a journey if appropriate.
Explore related journeys or tell Peri what you're working through.